CN115358975A - Method for performing interpretable analysis on brain tumor segmentation deep learning network - Google Patents
Method for performing interpretable analysis on brain tumor segmentation deep learning network Download PDFInfo
- Publication number
- CN115358975A CN115358975A CN202210894707.6A CN202210894707A CN115358975A CN 115358975 A CN115358975 A CN 115358975A CN 202210894707 A CN202210894707 A CN 202210894707A CN 115358975 A CN115358975 A CN 115358975A
- Authority
- CN
- China
- Prior art keywords
- decision
- path
- population
- tree
- decision tree
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/0002—Inspection of images, e.g. flaw detection
- G06T7/0012—Biomedical image inspection
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/70—Determining position or orientation of objects or cameras
- G06T7/73—Determining position or orientation of objects or cameras using feature-based methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10072—Tomographic images
- G06T2207/10088—Magnetic resonance imaging [MRI]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20081—Training; Learning
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/20—Special algorithmic details
- G06T2207/20084—Artificial neural networks [ANN]
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
- G06T2207/30016—Brain
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/30—Subject of image; Context of image processing
- G06T2207/30004—Biomedical image processing
- G06T2207/30096—Tumor; Lesion
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Biophysics (AREA)
- Software Systems (AREA)
- Evolutionary Computation (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Medical Informatics (AREA)
- Nuclear Medicine, Radiotherapy & Molecular Imaging (AREA)
- Radiology & Medical Imaging (AREA)
- Quality & Reliability (AREA)
- Image Analysis (AREA)
Abstract
The invention discloses a method for implementing interpretable analysis on a deep learning segmentation network based on a nuclear magnetic resonance image, and belongs to the field of computer vision and pattern recognition. The method comprises the following specific steps: 1. performing segmentation calculation on the nuclear magnetic resonance image by using a segmentation network; 2. extracting a radiologic characteristic from the nuclear magnetic resonance image; 3. constructing a decision tree group for fitting the segmentation network by utilizing the radiologic characteristics and carrying out structural optimization on the decision tree group through an evolutionary algorithm; 4. extracting important decision paths from the tree group; 5. and analyzing the decision path set aiming at the target to be explained, and screening and excavating the core rules through the associated features. The method optimizes the decision tree set by using the evolutionary algorithm, improves the simulation quality of the deep learning network, optimizes the tree group structure and reduces the rule extraction difficulty; the provided key decision path extraction method and the associated feature screening method can explain decision logic of a deep learning network in a tumor boundary segmentation process.
Description
Technical Field
The invention belongs to the field of computer vision and pattern recognition, and particularly relates to an interpretability analysis method for a deep learning network based on multi-sequence Magnetic Resonance Imaging (MRI) data through a designed tree set model based on evolutionary optimization. The method can reduce the scale of the tree group, improve the segmentation precision of the tree group, is beneficial to acquiring a characteristic rule which is more concentrated relative to a random forest, extracts an important decision path under the condition of a plurality of trees by using the path screening algorithm on the basis, and realizes interpretability analysis of the brain tumor segmentation deep learning network on the basis of the acquired associated characteristics.
Background
Brain gliomas are one of the most common primary intracranial tumors that arise from the canceration of brain and spinal glial cells. In brain tumor segmentation tasks, deep learning networks are typically used for segmentation. Currently, artificial intelligence provides a more effective method for computer-aided diagnosis, making the diagnostic process more effective and available to the general public. However, the deep learning method lacks interpretability in the field of medical image processing, and the expression and application value of the deep learning method are questioned. Understanding the nature of the model and interpretation of model decisions is crucial in assisting physicians in making brain tumor diagnoses.
Currently existing machine learning interpretability methods can be divided into internal interpretability models and model-independent interpretability models. Model-independent interpretability methods often fall into two categories, one is based on sample interpretation, that is, the behavior of a model is interpreted by selecting a particular instance of a dataset. Another way to approximate the original black-box model is to use a traditional interpretable model, such as a decision tree as a proxy model to interpret decisions inside the model, but the decision tree as a proxy model is often affected by the accuracy of the decision tree. For better global agents, the ensemble learning model works better than the decision tree, but the large tree group makes the model difficult to interpret. Based on the problems, the evolutionary optimization tree set model is utilized, so that fewer decision paths can be obtained under the precision close to that of a random forest, and the rule extraction difficulty is reduced; meanwhile, an important decision path is obtained by designing a tree set model characteristic path extraction method, and the associated characteristics of the tumor subregion are obtained by an associated characteristic extraction method. The method can explain decision logic of the black box model in the boundary segmentation process, and provides a more promising approach for exploring and improving the interpretability of the machine learning model.
Disclosure of Invention
The method aims at the problems that a decision tree is adopted in interpretability analysis in a related research method to fit a deep learning model, the effect is poor, the model based on ensemble learning is good in effect, but the model is still difficult to understand due to the large tree group, and the like. The invention discloses a novel interpretability method for realizing a depth model by a tree set model based on a decision tree group. For this reason, key technologies to be solved include: and (3) evolution optimization tree set construction, important path extraction and associated feature acquisition.
In order to achieve the purpose, the specific technical scheme of the invention is as follows:
an interpretable analysis method for a brain tumor segmentation deep learning network based on multi-sequence Magnetic Resonance Imaging (MRI) data through designed decision tree group-based model fitting and analytical calculation, the method comprises the following steps:
step 1: data preparation, specifically:
MRI data for each patient is a multi-modality image that may contain multiple image sequences (e.g., T1ce, T2, and Flair), and typically a brain tumor lesion will contain multiple distinct sub-regions, such as a necrotic non-enhanced core region (NET), a peritumoral edema region (ED), an enhanced core region (ET);
step 2: the data preprocessing specifically comprises the following steps: preprocessing the image by using linear function normalization and bias field correction technology;
and step 3: generating a fitting model based on a decision tree group through an evolutionary optimization process, specifically:
step 3.1: the method comprises the steps of radiosomic feature extraction, specifically: firstly, setting the patch width of an acquired pixel point to be 10, recording the coordinate of each pixel point in a tumor subarea as (i, j), and expressing all pixel point sets P in the tumor subarea as:
P＝{(i 1 ,j 1 ),(i 2 ,j 2 ),...,(i n ,j n )}
traversing each pixel point in the pixel point set P, obtaining all characteristics of the current tumor subregion, including 42-dimensional first-order characteristics, 402-dimensional second-order characteristics and 444-dimensional characteristics, and then performing the operation on each tumor subregion to obtain a characteristic two-dimensional table F _ P, which can be expressed as:
F_P＝{(F i ,p j )|0≤i≤m,0≤j≤n,p j ∈P}
wherein, F i Represents a feature, p j Representing pixel points, i and j represent index values, m represents 444-dimensional features, n represents the number of the pixel points, and P represents a pixel point set;
step 3.2: initializing a decision tree population for fitting calculation, specifically:
step 3.2.1: taking the characteristic two-dimensional table F _ P as the input of a fitting model, performing fitting calculation on a deep learning segmentation network to be analyzed, and firstly, generating an initialized decision tree group T 0 It can be expressed as:
T 0 ＝{at 1 ,at 2 ,...,at m ,rt 1 ,rt 2 ,...,rt n ,st 1 ,st 2 ,...,st i }
wherein, at m Representing a decision tree generated using the entire set of features, m representing the number of at's in the population of base learners; rt is a group of animals n The decision tree is generated by randomly generating nodes without depending on a data set; n represents the number of rt in the group of basis learners; st i Randomly selecting partial features and a decision tree generated by partial data, wherein i represents the number of st in a base learner group, and outputting an initialized decision tree group with the group scale of m + n + i;
step 3.2.2: initializing an integrator group, specifically: generating a binary coding population according to the population of the base learners in the step 3.2.1, wherein the coding length of each individual is m + n + i, each gene locus is uniquely mapped to a sub-tree in the population of the base learners, the mapped sub-tree is selected to be added into an integrator determined by the binary coding when the gene locus is 1, and if the gene locus is 0, the opposite is true, using en _ len to represent the scale of the integrator population, and dt _ len to represent the number of decision trees contained in the individual integrators, so as to generate an initialized integrator population;
step 3.3: optimizing the fitting effect of the decision tree population on the depth segmentation network by using evolutionary search specifically comprises the following steps:
step 3.3.1: the searching process of the new group specifically comprises the following steps: firstly, a new sub-tree population is generated in the sub-tree population by using a tree-based coding cross mutation process, and then a new combined population is generated in the integrator population by using binary coding cross mutation operation. The process is a bidirectional evolution search process, a new decision tree group and a new tree group integrated group are synchronously generated, and the search process is stopped when the group scale reaches the upper limit;
step 3.3.2: evaluating the fitting effect of the integrator individuals, specifically: performing multi-target evaluation by taking the segmentation precision of the integrator individual on different tumor sub-regions and the number of decision trees contained in the integrator as two different evaluation indexes, reserving high-quality integrator individuals, and deleting disadvantaged individuals until the lower limit of the population quantity is met;
step 3.3.3: after the termination condition is reached, outputting a corresponding decision tree group according to the best integrator individual, otherwise, returning to the step 3.3.1;
and 4, step 4: analyzing the rules in the decision tree group output in the step 3.3.3, specifically:
step 4.1: extracting a characteristic path of the decision tree set model, specifically: in the decision tree group obtained in step 3.3.3, the decision paths that can correctly predict different tumor sub-regions in all decision trees in the tree group are searched, pixel blocks of all tumor sub-regions are circularly traversed, and decision path sets a of different tumor sub-regions are respectively obtained, which can be expressed as:
A＝{A 1 ,A 2 ,A 3 ,...,A N }
wherein, N represents the number of all decision paths in the path set;
step 4.2: the repeated path screening specifically comprises the following steps: inputting a path set A, comparing an ith (i =1,2,3.. Cndot., N) path in the path set A with an ith +1 path in the path set, and deleting a duplicate path if the paths are equal; outputting the path set A until all paths in the path set A are not repeated;
step 4.3: and (3) performing correlation characteristic screening on a target tumor subarea needing to be explained, wherein the correlation characteristic screening method specifically comprises the following steps:
step 4.3.1: the characteristics of one decision path in the path set a are represented as follows:
wherein, i represents the path in the path set, N represents the number of the paths in the path set, j represents the characteristic of the ith path, and M represents the characteristic number of the ith path;
step 4.3.2: comparing two paths in the path set, wherein the specific method comprises the following steps: taking any two different paths in the path set A as input, and comparing the two paths in sequence
And with
Whether equal, the two paths can be expressed as:
wherein p and q are 1, i.e. path A 1 、A 2 The first characteristic is that step 4.3.3 is executed, and the process is transferred to step 4.3.5 until all paths are compared;
step 4.3.3: if it is
Go to step 4.3.4 if
And q ≠ M, q self-increments by 1 and repeats the procedure if
q = M and p ≠ M, making q 1 and p self-increment 1 and repeat the step, if q = M and p = M, go to step 4.3.2;
step 4.3.4: recording the current p and q as p 'and q', and enabling the p 'and the q' to be simultaneously and automatically increased by 1 until the current p and the current q are the p 'and the q' are recorded
Record the current associated decision path as
If the SF is not equal to any stored continuous subset of the associated decision path, storing the SF, if the stored associated decision path is not equal to any stored continuous subset of the SF, deleting the associated decision path, and turning to step 4.3.3;
step 4.3.5: and outputting all the stored associated decision characteristics to generate a decision rule set.
Drawings
FIG. 1 is a flow chart of an interpretability analysis of a decision tree group-based tree set model according to an embodiment of the present invention.
FIG. 2 is a diagram illustrating the composition of feature vectors in an example of the present invention.
FIG. 3 is a schematic diagram of generation of an optimal evolution integration model in an embodiment of the present invention.
FIG. 4 is a diagram illustrating the extraction of important decision paths of the tree set model according to the embodiment of the present invention.
FIG. 5 is a diagram illustrating the result of extracting important features according to an embodiment of the present invention.
Detailed description of the preferred embodiment
The following describes a specific implementation method for a U-Net network by taking Net and ED area division as an example with reference to the accompanying drawings. The following examples are only for illustrating the technical solutions of the present invention more clearly, and the protection scope of the present invention is not limited thereby. FIG. 1 shows a process of an interpretable analysis method of a tree set model based on a decision tree group, which includes the following steps:
step 1: inputting four original MRI sequence images, and training a U-Net network to obtain a segmentation result graph;
step 2: carrying out image omics feature extraction on NET and ED areas by using the U-Net network segmentation result and 4 sequence images in MRI data to form a feature two-dimensional table F _ P, wherein the feature vector result of each pixel point refers to FIG. 2;
and 3, step 3: referring to fig. 3, an initialized decision tree population and an integrator population are generated based on the feature two-dimensional table F _ P as an input, and then a fitting model is optimized by using an evolutionary algorithm, and a corresponding decision tree population is output according to the best integrator individual;
and 3, step 3: performing interpretable analysis on the decision tree group model of the fitted U-Net network, specifically comprising a step 4 and a step 5;
and 4, step 4: referring to fig. 4, a decision tree set model for evolutionary optimization is input, and a path extraction algorithm of the decision tree set model is used to perform path extraction on ED and NET regions respectively;
and 5: inputting the extracted decision path, and using a correlation feature screening algorithm to obtain important correlation paths in the ED and NET regions, respectively, with the result referring to fig. 5.
Claims (1)
1. An interpretable analysis method for a brain tumor segmentation deep learning network based on multi-sequence Magnetic Resonance Imaging (MRI) data through designed decision tree group-based model fitting and analytic calculation, the method comprising the following steps:
step 1: data preparation, specifically:
MRI data for each patient is a multi-modality image that may contain multiple image sequences (e.g., T1ce, T2, and Flair), and typically a brain tumor lesion will contain multiple distinct sub-regions, such as a necrotic non-enhanced core region (NET), a peritumoral edema region (ED), an enhanced core region (ET);
and 2, step: the data preprocessing specifically comprises the following steps: preprocessing an image by using a linear function normalization and bias field correction technology;
and 3, step 3: generating a fitting model based on a decision tree group through an evolutionary optimization process, which specifically comprises the following steps:
step 3.1: the method comprises the steps of radiosomic feature extraction, specifically: firstly, setting the patch width of the acquired pixel points to be 10, and noting that the coordinate of each pixel point in the tumor subregion is (i, j), all the pixel point sets P in the tumor subregion can be expressed as:
P＝{(i 1 ,j 1 ),(i 2 ,j 2 ),…,(i n ,j n )}
traversing each pixel point in the pixel point set P, obtaining all characteristics of the current tumor subregion, including 42-dimensional first-order characteristics, 402-dimensional second-order characteristics and 444-dimensional characteristics, and then performing the operation on each tumor subregion to obtain a characteristic two-dimensional table F _ P, which can be expressed as:
F_P＝{(F i ,p j )|0≤i≤m,0≤j≤n,p j ∈P}
wherein, F i Represents a feature, p j Expressing pixel points, i and j express index values, m expresses 444-dimensional characteristics, n expresses the number of the pixel points, and P expresses a pixel point set;
step 3.2: initializing a decision tree population for fitting calculation, specifically:
step 3.2.1: taking the characteristic two-dimensional table F _ P as the input of a fitting model, performing fitting calculation on a deep learning segmentation network to be analyzed, and firstly, generating an initialized decision tree group T 0 It can be expressed as:
T 0 ＝{at 1 ,at 2 ,…,at m ,rt 1 ,rt 2 ,…,rt n ,st 1 ,st 2 ,…,st i }
therein, at m Representing a decision tree generated using the entire set of features, m representing the number of at's in the population of base learners; rt is a group of animals n The decision tree is generated by randomly generating nodes without depending on a data set; n represents the number of rt in the population of base learners; st i Randomly selecting partial features, generating a decision tree by partial data, wherein i represents the number of st in a population of a base learner, and outputting an initialized decision tree population with the population scale of m + n + i;
step 3.2.2: initializing an integrator group, specifically: generating a binary coding population according to the base learner population of the step 3.2.1, wherein the individual codes have lengths of m + n + i, each gene position is uniquely mapped to a sub-tree in the base learner population, the mapped sub-tree is selected to be added to an integrator determined by the binary coding when the gene position is 1, if the gene position is 0, the scale of the integrator population is represented by en _ len, and the number of decision trees contained in the integrator individual is represented by dt _ len, so that an initialization integrator population is generated;
step 3.3: the method for optimizing the fitting effect of the decision tree population on the depth segmentation network by using evolutionary search specifically comprises the following steps:
step 3.3.1: the searching process of the new group specifically comprises the following steps: firstly, a new sub-tree population is generated in the sub-tree population by using a tree-based coding cross mutation process, and then a new combined population is generated in the integrator population by using binary coding cross mutation operation. The process is a bidirectional evolutionary search process, a new decision tree group and a new tree group integrated group are synchronously generated, and the search process is stopped when the group scale reaches the upper limit;
step 3.3.2: evaluating the fitting effect of the integrator individuals, specifically: performing multi-target evaluation by taking the segmentation precision of the integrator individual on different tumor sub-regions and the number of decision trees contained in the integrator as two different evaluation indexes, reserving high-quality integrator individuals, and deleting disadvantaged individuals until the lower limit of the population quantity is met;
step 3.3.3: after the termination condition is reached, outputting a corresponding decision tree group according to the best integrator individual, otherwise, returning to the step 3.3.1;
and 4, step 4: analyzing the rules in the decision tree group output in the step 3.3.3, specifically:
step 4.1: extracting a characteristic path of the decision tree set model, specifically: in the decision tree group obtained in step 3.3.3, the decision paths that can correctly predict different tumor sub-regions in all decision trees in the tree group are searched, pixel blocks of all tumor sub-regions are circularly traversed, and decision path sets a of different tumor sub-regions are respectively obtained, which can be expressed as:
A＝{A 1 ,A 2 ,A 3 ,…,A N }
wherein, N represents the number of all decision paths in the path set;
step 4.2: the repeated path screening specifically comprises the following steps: inputting a path set A, starting from the ith (i =1,2,3.., N) path in the path set A, comparing with the (i + 1) th path in the path set, and deleting the repeated path if the paths are equal; outputting the path set A until all paths in the path set A are not repeated;
step 4.3: and (3) performing correlation characteristic screening on a target tumor subarea needing to be explained, wherein the correlation characteristic screening method specifically comprises the following steps:
step 4.3.1: the characteristics of one decision path in the path set a are represented as follows:
wherein, i represents the path in the path set, N represents the number of the paths in the path set, j represents the characteristic of the ith path, and M represents the characteristic number of the ith path;
step 4.3.2: comparing two paths in the path set, wherein the specific method comprises the following steps: taking any two different paths in the path set A as input and comparing the paths in sequence
And with
Whether equal, the two paths can be expressed as:
wherein p and q are 1, i.e., path A 1 、A 2 The first characteristic is that step 4.3.3 is executed until all paths are compared, and the step 4.3.5 is carried out;
step 4.3.3: if it is
Go to step 4.3.4 if
And q ≠ M, q is increased by 1 and the procedure is repeated if
q = M and p ≠ M, making q 1 and p self-increment 1 and repeat the step, if q = M and p = M, go to step 4.3.2;
step 4.3.4: recording the current p and q as p 'and q', and enabling p 'and q' to increase by 1 at the same time until the current p and q are the p 'and q' respectively
Record the current associated decision path as
If the SF is not equal to any stored continuous subset of the associated decision path, storing the SF, if the stored associated decision path is not equal to any stored continuous subset of the SF, deleting the associated decision path, and turning to step 4.3.3;
step 4.3.5: and outputting all the stored associated decision characteristics to generate a decision rule set.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210894707.6A CN115358975A (en) | 2022-07-28 | 2022-07-28 | Method for performing interpretable analysis on brain tumor segmentation deep learning network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210894707.6A CN115358975A (en) | 2022-07-28 | 2022-07-28 | Method for performing interpretable analysis on brain tumor segmentation deep learning network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115358975A true CN115358975A (en) | 2022-11-18 |
Family
ID=84031056
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210894707.6A Pending CN115358975A (en) | 2022-07-28 | 2022-07-28 | Method for performing interpretable analysis on brain tumor segmentation deep learning network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115358975A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116704208A (en) * | 2023-08-04 | 2023-09-05 | 南京理工大学 | Local interpretable method based on characteristic relation |
-
2022
- 2022-07-28 CN CN202210894707.6A patent/CN115358975A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116704208A (en) * | 2023-08-04 | 2023-09-05 | 南京理工大学 | Local interpretable method based on characteristic relation |
| CN116704208B (en) * | 2023-08-04 | 2023-10-20 | 南京理工大学 | Local interpretable method based on characteristic relation |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111476292B (en) | Small sample element learning training method for medical image classification processing artificial intelligence | |
| Khan et al. | Transfer learning with intelligent training data selection for prediction of Alzheimer’s disease | |
| Li et al. | 3D neuron reconstruction in tangled neuronal image with deep networks | |
| Zou et al. | Deep learning based feature selection for remote sensing scene classification | |
| CN113314205B (en) | Efficient medical image labeling and learning system | |
| CN107133496B (en) | Gene feature extraction method based on manifold learning and closed-loop deep convolution double-network model | |
| CN111242948B (en) | Image processing method, image processing device, model training method, model training device, image processing equipment and storage medium | |
| CN108304573A (en) | Target retrieval method based on convolutional neural networks and supervision core Hash | |
| Bergstra et al. | Making a science of model search | |
| Sumbul et al. | Informative and representative triplet selection for multilabel remote sensing image retrieval | |
| CN115098620A (en) | Cross-modal Hash retrieval method for attention similarity migration | |
| CN115661165A (en) | Glioma fusion segmentation system and method based on attention enhancement coding and decoding network | |
| CN115358975A (en) | Method for performing interpretable analysis on brain tumor segmentation deep learning network | |
| CN114241267A (en) | Structural entropy sampling-based multi-target architecture search osteoporosis image identification method | |
| Song et al. | Semi-MapGen: translation of remote sensing image into map via semisupervised adversarial learning | |
| CN116612339B (en) | Construction device and grading device of nuclear cataract image grading model | |
| CN116431004B (en) | Control method and system for interactive behavior of rehabilitation robot | |
| CN113208641A (en) | Pulmonary nodule auxiliary diagnosis method based on three-dimensional multi-resolution attention capsule network | |
| CN116958693A (en) | Image analysis method, apparatus, device, storage medium, and program product | |
| CN116386803A (en) | Cytopathology report generation method based on graph | |
| CN116188428A (en) | Bridging multi-source domain self-adaptive cross-domain histopathological image recognition method | |
| Cao et al. | Understanding 3D point cloud deep neural networks by visualization techniques | |
| Yu et al. | AI in Paleontology | |
| CN112307288A (en) | User clustering method for multiple channels | |
| Cheng | The cross-field DBN for image recognition |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |